Investigate support for streaming mode

[HennerM]

Previously reported in #2 (comment) by @aamir-s18

Streaming mode could be useful for very big models. It can help in real-time use cases where we can improve User Experience by generating one token at a time.

Triton Server supports streaming with decoupled models

It needs to be investigated how CTranslate can be used to get decoded tokens one-by-one. Additionally this might be trickier in a beam decode setting, unless we are willing to always return the best guess which could flip previous words

[aamir-s18]

I think this here is relevant for this issue.